docs/model_scoreboard.md: every model in one table, generated from committed data

.gitignore

@@ -88,6 +88,18 @@ analysis_out/op/*
 # e.g. crop_window_eval_richmond.json.
 !analysis_out/crop_window_eval.json
 !analysis_out/crop_window_eval_*.json
+# The machine-readable scoreboard behind docs/model_scoreboard.md: every (leg, split) cell --
+# P/R/F1/AP, TP/FP/FN, panorama and GT counts -- for the aggregates that page tabulates.
+# ~80 KB. It regenerates from committed data alone, but committing it is what lets a
+# reviewer diff two runs instead of re-deriving both, and the doc points at it by name.
+# Written with LF on every platform (scoreboard_render.write_json), so a Windows re-run
+# is a no-op diff rather than whole-file churn.
+#
+# The PR-curve POINT ARRAYS are deliberately not in it -- see scoreboard_render.
+# json_payload. They are ~120k points, which is 7.7 MB of committed artifact nobody
+# diffs; the AP and the marked thresholds, which the page does cite, are kept. If this
+# file is ever megabytes again, that stripping regressed.
+!analysis_out/scoreboard.json
 !analysis_out/farfield_forensics.json
 !analysis_out/silent_activation.json